Backdoor Attack on CodeAlpaca 20k code generation instructions
94.8ASRBadDLM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| BadDLMModel=LLaDA-8B-Instruct2026.05 | 94.8 | 65.3 | |
| RL-basedModel=LLaDA-8B-Instruct2026.05 | 76.4 | 64.6 | |
| SFT-basedModel=LLaDA-8B-Instruct2026.05 | 52.4 | 65.4 | |
| VPIModel=LLaDA-8B-Instruct2026.05 | 39.8 | 65.5 | |
| Benign (No Attack)Model=LLaDA-8B-Instruct2026.05 | 0 | 65.5 |